Overview
Sequential equilibrium is a solution concept for extensive-form games—games represented as trees with moves, information sets, and payoffs—when players may have private information or face uncertainty about others. It refines Nash equilibrium by requiring that strategies be optimal at every point in the game given beliefs, and that beliefs about what has happened are consistent with the strategies via Bayes’ rule wherever possible. In plain terms: a sequential equilibrium specifies (i) how players plan to act at every decision point and (ii) what they would believe if any decision point were reached, including those off the expected path; and it demands that these plans and beliefs fit together coherently.
Sequential equilibrium matters because it filters out non-credible Nash equilibria that rely on implausible threats or arbitrary beliefs about off-path play. It is the workhorse for analyzing signaling, reputation, bargaining, and entry deterrence—core topics for strategy, regulation, and market design.
Origins and Credit
David Kreps and Robert Wilson introduced sequential equilibrium in 1982. Their goal was to formalize “perfect Bayesian” reasoning with a rigorous notion of belief consistency. They showed that sequential equilibria exist in finite extensive-form games and developed practical tools based on trembles—small probabilities of every move—to discipline off-path beliefs.
Core Idea and Mechanics
An extensive-form game has players, possible histories of play, information sets (collections of histories among which the player moving cannot distinguish), and payoffs. A strategy specifies an action for a player at every information set where they might move. An assessment is a pair (σ, μ): σ is a profile of strategies (one for each player), and μ is a belief system giving, for each information set, probabilities over which histories in that set have occurred.
Sequential equilibrium requires two conditions:
- Sequential rationality: Given beliefs μ about how the game has unfolded, each player’s strategy σ maximizes expected payoff at every information set they control. This means there are no profitable deviations even at information sets that are reached with probability zero under σ.
- Consistency: Beliefs are derived from strategies via Bayes’ rule on-path (where the information set is reached with positive probability). Off-path, beliefs must be limits of Bayes-consistent beliefs from sequences of fully mixed “trembling” strategies that approximate σ. Intuition: imagine small mistakes make every action possible with tiny probability; compute beliefs accordingly; then let the mistake probabilities go to zero.
This approach pins down how players should interpret surprising moves and rules out equilibria that depend on arbitrary or conspiratorial off-path beliefs.
Relationship to other concepts:
- Nash equilibrium: Every sequential equilibrium is a Nash equilibrium, but not vice versa.
- Subgame perfect equilibrium (SPE): In games with perfect information (no nontrivial information sets), sequential equilibrium coincides with SPE. With imperfect information, sequential equilibrium refines SPE by adding belief consistency.
- Perfect Bayesian equilibrium (PBE): PBE also pairs strategies and beliefs with sequential rationality, but it is less specific about how off-path beliefs are formed. Sequential equilibrium imposes a particular discipline via trembles, making it a more tightly defined refinement.
Key Assumptions and Conditions
- Finite extensive-form structure and perfect recall: Players remember their past actions and information. Existence results rely on finiteness and, often, perfect recall.
- Bayesian updating: Beliefs are updated via Bayes’ rule where defined.
- Expected-utility maximization and common knowledge of the game: Players best respond given beliefs; the rules and payoffs are common knowledge.
- Trembles for consistency: Off-path beliefs must be limits of beliefs induced by fully mixed strategy profiles.
Implications
- Credible dynamic behavior: Eliminates equilibria sustained by empty threats; strategies must be optimal at every conceivable decision point.
- Disciplined off-path beliefs: Off-path interpretations cannot be arbitrary; they must be justifiable as limits of vanishing mistakes.
- Refined predictions in signaling and reputation: Helps determine which separating or pooling outcomes are plausible when actions convey information (e.g., pricing as a signal of quality).
- Policy and design relevance: Useful in auction design, contract renegotiation, and regulatory processes where unexpected moves occur and must be interpreted.
Example in Practice
Warranty as a signal of product quality. A manufacturer privately knows whether its product is high or low quality. Stage 1: the firm offers either a generous warranty (W) or a limited warranty (L). Stage 2: observing the warranty but not true quality, consumers decide to purchase or not. Payoffs reflect that high quality is less costly to warrant (fewer claims). Consumers’ value is higher for high quality.
A plausible separating sequential equilibrium is: high-quality types choose W; low-quality types choose L; consumers buy if they see W and decline if they see L. Beliefs μ assign high probability to high quality upon observing W and to low quality upon observing L. Sequential rationality holds because (i) given these beliefs, consumers’ buy/decline choices are payoff-maximizing, and (ii) given consumers’ responses, the high-quality firm prefers W (its net profits are higher with sales and manageable claim costs) while the low-quality firm prefers L (a generous warranty would attract buyers but expected claims would erase profits).
Consistency requires that if, contrary to equilibrium, a low-quality firm chose W or a high-quality firm chose L, consumers’ beliefs about quality in those off-path cases arise as limits from small trembles (e.g., a tiny chance that any type accidentally picks the “wrong” warranty). This disciplines off-path beliefs and ensures the separating outcome does not rely on ad hoc pessimism or optimism.
Managerial lesson: if a costly action is less burdensome for high types than low types (a “single-crossing” cost structure), then warranties, return policies, or service levels can credibly signal quality. Sequential equilibrium shows when such signals are self-enforcing and how customers should update from unexpected offers.
Limitations and Common Misunderstandings
- Multiplicity remains: Games often have multiple sequential equilibria (e.g., pooling and separating in signaling). Selection may require additional refinements (intuitive criterion, D1, divinity) or empirical grounding.
- Not “Bayes everywhere”: Bayes’ rule only pins down beliefs on the path of play. Off-path beliefs are disciplined via trembles, not uniquely determined.
- Different from PBE: Perfect Bayesian equilibrium is a family of conditions; some PBEs are not sequential equilibria because their off-path beliefs cannot be justified as limits of trembles.
- Modeling detail matters: How information sets are drawn or whether players have perfect recall affects predictions. Coarse or mis-specified information can artificially permit or rule out equilibria.
- Computational complexity: In large games, verifying sequential rationality at all information sets and constructing consistent beliefs can be demanding.
Related Concepts (names only)
Nash Equilibrium; Subgame Perfect Equilibrium; Perfect Bayesian Equilibrium; Trembling-Hand Perfection; Signaling Games; Reputation Effects.
